In this method, linguistic phenomena are investigated that consist of explicit representations of facts about language via precise and well-understood knowledge representations (Laporte, 2005). Symbolic systems evolve from human-developed rules and lexicons, which generate the relevant information on which this approach is based. Once the rules have been defined and identified, a document is analyzed to pinpoint the exact conditions, which validates them. All such rules associated with semantic objects generate networks describing their hierarchical structure. In fact, highly associated concepts exhibit directly linked properties, whereas moderately or weakly related concepts are linked through other semantic objects. Symbolic methods have been widely exploited in a variety of research contexts such as information extraction, text categorization, ambiguity resolution, explanation-based learning, decision trees, and conceptual clustering.

This approach is based on observable data and large documents to develop generalized models based on smaller knowledge datasets and significant linguistic or world knowledge (Manning and Schütze, 1999). The set of states is associated with probabilities. Several techniques can be used to investigate their structures, such as the hidden Markov model, in which the set of status is regarded as not directly observable. The techniques have many applications such as parsing rule analysis, statistical grammar learning, and statistical machine translation, to name but a few.

This approach combines statistical learning with representation techniques to allow an integration of statistical tools with logic-based rule manipulation, generating a network of interconnected simple processing units (often associated with concepts) with edge weights representing knowledge. This typically creates a rich system with an interesting dynamical global behavior induced by the semantic propagation rules. In Troussov et al. (2010), a connectionist distributed model is investigated pointing toward a dynamical generalization of syntactic parsing, limited domain translation tasks, and associative retrieval.

The most basic level in an NLP system is based on lexical analysis, which deals with words regarded as the atomic structure of text documents; in particular, it is the process that takes place when the basic components of a text are analyzed and grouped into tokens, which are sequences of characters with a collective meaning (Dale et al., 2000). In other words, lexical analysis techniques identify the meaning of individual words, which are assigned to a single part-of-speech (POS) tag. Lexical analysis may require a lexicon, which is usually determined by the particular approach used in a suitably defined NLP system as well as the nature and extent of information inherent in the lexicon. Mainly, lexicons may vary in terms of their complexity because they can contain information about the semantic information related to a word. Moreover accurate and comprehensive sub-categorization lexicons are extremely important for the development of parsing technology as well as for any NLP application that relies on the structure of information related to predicate-argument structure. More research is currently being carried out to provide better tools for analyzing words in semantic contexts (see Korhonen et al., 2006 for an overview).

POS tagging is one of the first steps in text analysis because it allows us to attach a specific syntactic definition (noun, verb, adjective, etc.) to the words that are part of a sentence. This task tends to be accurate because it relies on a set of rules that are usually unambiguously defined. Consider the word “book.” Depending on the sentence to which it belongs, it might be a verb or a noun. Consider “a book on a chair” and “I will book a table at the restaurant.” The presence of specific keywords such as “a” in the former and “I will” in the latter provides important clues as to the syntactic role that “book” has in the two sentences. One of the main reasons for the overall accuracy of POS tagging is that a semantic analysis is often not required. It is based only on the position of the word and its neighboring words.

Once the POS tagging of a sentence has identified the syntactic roles of each word, we can consider such a sentence in its entirety. The main difference with POS tagging is that parsing enables the identification of the hierarchical syntactic structure of a sentence. Consider, for example, the parsing tree structure of the sentence “This is a parsing tree,” depicted in Figure 11.3. Each word is associated with a POS symbol that corresponds to its syntactic role (Manning and Schütze, 1999).

An important aspect of text analysis is the ability to determine the type of the entities within a text fragment. More specifically, determining whether a noun refers to a person, an organization, or a geographical location (to name but a few) substantially contributes to the extraction of accurate information and provides the tools for a deeper understanding. For example, the analysis of “Dogs and cats are the most popular pets in the UK” would identify that dogs and cats are animals and the UK is a country. Clearly, there are many instances where this depends on the context. Think of “India lives in Manchester.” Anyone reading such a sentence would interpret India as the name of a specific person, and rightly so. However, a computer might not be able to do so and may decide that it is a country. We know that a country would not be able to “live” in a city. It is just common sense. Unfortunately, computers do not have the ability to discern what common sense is. They might be able to guess according to the structure of a sentence or the presence of specific keywords. Nevertheless, it is precisely that—a guess. This is an effective example of semantic understanding, which comes naturally to humans but is a complex, if impossible, task for computers.

Co-reference resolution is the task of determining, which words refer to the same objects. An example is anaphora resolution, which is specifically concerned with the nouns or names to which words refer. Another instance of co-reference resolution is relation resolution, which attempts to identify to which individual entities or objects a relation refers. Consider the following sentence: “We are looking for a corrupted member of the panel.” Here, we are not looking for just any corrupted member of the panel, but for a specific individual.

The identification of relations among different entities within a text provides useful information that can be used to disambiguate subcomponents of the text fragment as well as determine quantitative and qualitative information linking such entities. For example, consider the sentence “Smoking potentially causes lung cancer.” Here, the act of smoking is linked to lung cancer by a causal relationship.

A crucial task in information extraction from textual sources is concept identification, which is typically defined as one or more keywords or textual definitions. The two main approaches in this task are supervised and unsupervised concept identification, depending on the level of human intervention in the system.

This procedure attempts to identify the general topic of a text by grouping a set of keywords that appear frequently in the documents. These are then associated with one or more concepts to determine the general concept trend.

This is a particular example of information extraction from text, which focuses on identifying trends of opinions across a population in social media. Broadly speaking, its aim is to determine the polarity of a given text, which identifies whether the opinion expressed is positive, negative, or neutral. This includes emotional states such as anger, sadness, and happiness, as well as intent: for example, planning and researching. Sentiment analysis can be an important tool in obtaining insight into criminal activity because it can detect a general state of mind shared by a group of individuals via their blog entries and social network discussions. A variety of tools are used in crime detection as well as prosecution that can automate law enforcement across several social media, including Web sites, e-mails, personal blogs, and video uploaded onto social platforms (Dale et al., 2000).

Semantic analysis determines the possible meanings of a sentence by investigating interactions among word-level meanings in the sentence. This approach can also incorporate the semantic disambiguation of words with multiple senses. Semantic disambiguation allows selection of the sense of ambiguous words, so that they can be included in the appropriate semantic representation of the sentence (Wilks and Stevenson, 1998). This is particularly relevant in any information retrieval and processing system based on ambiguous and partially known knowledge. Disambiguation techniques usually require specific information about the frequency with which each sense occurs in a particular document, as well as an analysis of the local context and the use of pragmatic knowledge of the domain of the document. An interesting aspect of this research field is concerned with the purposeful use of language in which the use of a context within the text is exploited to explain how extra meaning is part of some documents without actually being constructed in them. Clearly this is still being developed, because it requires an incredibly wide knowledge dealing with intentions, plans, and objectives (Manning and Schütze, 1999). Extremely useful applications in NLP can be seen in inferencing techniques, in which extra information derived from a wider context successfully addresses statistical properties (Kuipers, 1984).

The automatic translation of a text from one language into another one provides users with the ability to read texts quickly and fairly accurately in a variety of languages. This is a complex task involving several challenges, from mapping different syntactic structures to semantic interpretation and disambiguation.

Bayesian networks (BNs) are graphical models that capture independence relationships among random variables, based on a basic law of probability called Bayes’ rule (Pearl, 1998). They are a popular modeling framework in risk and decision analysis and have been used in a variety of applications such as safety assessment of nuclear power plants, risk evaluation of a supply chain, and medical decision support tools (Nielsen and Verner Jensen, 2009). More specifically, BNs are composed of a graph whose nodes represent objects based on a level of uncertainty, also called random variables, and whose edges indicate interdependence among them. In addition to this graphical representation, BNs contain quantitative information that represents a factorization of the joint probability distribution of all variables in the network. In fact, each node has an associated conditional probability table that captures the probability distribution associated with the node conditional for each possibility.

Nodes in BNs, which are connected by edges, imply that the corresponding random variables are dependent. Such dependence relations must therefore be extracted from textual information when they are present. The conditional dependencies in a BN are often based on known statistical and computational techniques. One of their strengths is the combination of methods from graph theory, probability theory, computer science, and statistics. Linguistically speaking, a dependence relation contains specific keywords that indicate that two concepts are related to a certain degree. Consider the sentence “Lung cancer is more common among smokers.” There is little doubt that we would interpret this as a clear relation linking lung cancer with smoking. However, there is not a precise linguistic definition to determine a relationship between two concepts from the text. The reason is that it depends on the context.

Mapping a representative to a specific variable is closely linked to the task of relations extraction. However, this is partially a modeling choice by the user based on the set of concepts in which she or he is interested. Consider again the sentence “Smoking causes lung cancer.” If this were rephrased as “Smokers are more likely to develop lung cancer,” we would need to ensure that “smoking” and “smokers” are identified as a single variable associated with the act of smoking. In a variety of cases, this can be addressed by considering synonymy. However, as in our example, it might also happen that they refer to the same concept, rather than being the same concept. FCA is a computational technique that can be successfully applied in this particular context.

This step aggregates the information gathered in the previous two steps to output the structure of the BN. This includes evaluating the credibility of each dependence relation, determining whether the dependence stated is direct or indirect, and ensuring that no cycles are created in the process of orienting the edges.

This step involves processing the textual sources to extract information about the probability of variables. This includes the search for both numerical information and quantitative statements such as “Smoking increases the chances of lung cancer” and “Nonsmokers are less likely to get lung cancer.”

This step obtains a fully defined BN network automatically, typically with the help of the user. This also includes state identification and combining conflicting and partial information in a rigorous way.

The general architecture of the extraction of BNs from social media for detecting and assessing criminal activity typically consists of the following components (see also Figure 11.8):

Automatic extraction of relevant information to define and populate BNs has a variety of applications in crime detection and management. Furthermore, the huge potential offered by the investigation of Big Data can provide critical insight into criminal activities. However, the extraction, assessment, and management of datasets that are continuously created poses challenges that have to be addressed to provide a valuable platform.